A causal-aware active learning method and system for catalyst discovery

Abstract

This invention belongs to the field of machine learning technology and discloses a causal awareness-based active learning method and system for catalyst discovery, including: S1, data initialization and candidate pool construction; S2, Gaussian process substitution model training; S3, causal graph learning and edge confidence estimation; S4, causal intervention value calculation; S5, expectation improvement calculation; S6, adaptive weight calculation; S7, joint acquisition function construction; S8, candidate selection and label acquisition; S9, iteration until convergence: repeating steps S2 to S8 until the maximum number of iterations is reached, outputting the final dataset and the corresponding causal graph. This invention not only optimizes prediction performance during automated iterative experiments but also clarifies the iterative causal structure of catalyst discovery and improves the quality of causal graph structure recovery during the iterative process.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, specifically relating to a causal perception-based active learning method and system for catalyst discovery. Background Technology

[0002] Catalyst discovery is a core R&D step in clean energy technologies such as water electrolysis, fuel cells, and carbon dioxide reduction. Traditional trial-and-error experimental methods are time-consuming and costly. In recent years, active learning has become an important paradigm for accelerating catalyst discovery. Standard active learning methods (such as uncertainty sampling and expectation improvement (EI)) can improve the information gain per experiment by iteratively selecting experiments with high information content.

[0003] However, existing active learning methods have a fundamental limitation: they only focus on reducing prediction uncertainty and do not explicitly consider the causal structure behind the data, thus failing to answer the mechanistic question of "why some catalysts perform better." Understanding causality is crucial in catalyst discovery—researchers need not only to know which descriptors predict catalytic activity, but also to identify which descriptors truly "lead" to high activity. For example, specific surface area may be highly correlated with overpotential, but this correlation may be driven by confounding factors such as particle size rather than a direct causal effect. Existing methods cannot distinguish between such causal descriptors and spurious descriptors.

[0004] In the prior art, patent document US20220245508A1 discloses an active learning system using causal network feedback, but its causal network correction relies on feedback from human domain experts, failing to achieve fully automatic causal structure updates. Patent document CN119067028A discloses a process optimization method combining causal inference and active learning, but its active learning only employs a single uncertainty sampling strategy, without explicitly jointly modeling the causal structure information gain and performance prediction optimization. While existing academic research touches upon causal active learning theories, these are all general frameworks and have not been customized for the high-dimensional descriptor space, DFT computational cost, and multiple types of descriptor features in catalyst discovery scenarios.

[0005] Furthermore, Chinese patents CN202210816626.4 and CN122046889A respectively use Gaussian process surrogate models for shield tunnel risk assessment and photolithography process optimization, but neither involves causal structure learning. Therefore, there is an urgent need for a fully automated catalyst discovery method and system that simultaneously optimizes predictive performance and clarifies causal structures during iterative experiments. Summary of the Invention

[0006] The purpose of this invention is to provide a causal perception-based active learning method for catalyst discovery, in order to solve the technical problem that existing technologies cannot clarify the iterative causal structure of catalyst discovery and optimize prediction performance during automated iterative experiments.

[0007] The aforementioned causal sensing active learning method for catalyst discovery includes the following steps: S1. Data initialization and candidate pool construction: Obtain the candidate catalyst pool and the initial catalyst label dataset; S2. Training of Gaussian process substitution model: Train the Gaussian process substitution model based on the current dataset to obtain the predicted mean and predicted variance of catalytic performance; S3. Causal Graph Learning and Edge Confidence Estimation: The PC algorithm is used to learn the causal structure of the current dataset to obtain the causal graph and the edge confidence of each edge in the causal graph. S4. Causal intervention value calculation: Based on the confidence of the updated front side, the confidence of the updated back side obtained by using the prediction mean of the Gaussian process replacement model as a pseudo-observation, and the weight of the uncertain side, the causal intervention value of each candidate sample is calculated. S5. Expected Improvement Calculation: Based on the predicted mean and predicted standard deviation of the Gaussian process substitution model, and the current best observation, calculate the expected improvement relative to the current best observation by minimizing the objective. S6. Adaptive weight calculation: Calculate adaptive weights based on the current iteration number and the maximum iteration number; S7. Construction of Joint Acquisition Function: Based on the adaptive weights, the causal intervention value and the expected improvement are linearly combined into a joint acquisition function; S8. Candidate Selection and Label Acquisition: Select the optimal candidate sample and obtain the true label based on the joint acquisition function; S9. Iterate until convergence: Repeat steps S2 to S8 until the maximum number of iterations is reached, and output the final dataset and the corresponding causal graph.

[0008] Preferably, in step S2, the Gaussian process substitution model uses a radial basis function kernel, and the hyperparameters are optimized by maximizing the marginal likelihood; when the dataset size is small, a fixed default length scale is used to avoid unstable hyperparameter estimation; when the amount of data is sufficient, the kernel hyperparameters are optimized by maximizing the marginal likelihood.

[0009] Preferably, in step S3, the causal graph is represented as a partially directed acyclic graph. In each iteration, the edge confidence of each edge in the causal graph is obtained through Bootstrap resampling. The edge confidence is obtained by statistically analyzing the frequency and directional consistency of each edge in the subsample causal graph. For each possible edge (i,j), the proportion of edges identified as existing and having the same direction in the subsample is calculated to obtain the edge confidence before the update. In a CPDAG, edges with undetermined directions but adjacency relationships are denoted as structurally uncertain edges.

[0010] Preferably, in step S4, a Gaussian process is used to replace the predicted mean of the model. As a pseudo-observation label, the simulation will Add the current dataset to form a temporary dataset, and rerun the Bootstrap PC algorithm on this temporary dataset to obtain the updated confidence scores. ; Calculate the change in edge confidence The formula is Calculate the uncertainty weights The formula is The causal intervention value is obtained by weighting and summing the uncertainty weights of each uncertain edge with the corresponding edge confidence changes. It is used to measure the potential clarifying effect of candidate samples on structurally uncertain edges.

[0011] Preferably, in step S5, based on the predicted mean and the predicted standard deviation Define the current best observation value as Standardized variables are Expected to be improved to Where Phi is the standard normal distribution function and phi is the standard normal probability density function; when the prediction standard deviation When =0, the expected improvement is Take 0.

[0012] Preferably, in step S6, the adaptive weight calculation formula is: Where N is the current iteration number, counted starting from 1. Maximum number of iterations; early stage Larger, prioritized causal exploration makes the joint acquisition function more biased towards IVs to clarify the causal structure; later Smaller values ​​are prioritized for prediction utilization, making the joint acquisition function more biased towards EI to select candidate samples that are more likely to improve target performance.

[0013] Preferably, in step S7, S7: the joint acquisition function is... ,in For adaptive weights, For the value of causal intervention, For improvement.

[0014] Preferably, in step S1, the candidate catalyst pool consists of candidate catalyst samples that have not yet obtained a true target catalytic performance label. The candidate samples are obtained from historical material database screening, composition space enumeration, structure generation rules and / or high-throughput pre-screening results, and each candidate sample has completed the catalyst descriptor vector calculation; the initial catalyst label dataset is formed by random sampling samples and / or preset seed samples in the candidate catalyst pool after density functional theory calculation or experimental labeling, and includes catalyst descriptor vectors and corresponding target catalytic performance labels.

[0015] Preferably, in step S2, the Gaussian process substitution model provides a predicted mean for the candidate sample x. and prediction variance Where k is the kernel function and K is the kernel matrix. I represents the variance of the observed noise, and I is the identity matrix.

[0016] This invention also provides a causal sensing active learning system for catalyst discovery, comprising: The data initialization module is used to obtain a candidate catalyst pool and an initial catalyst dataset. The candidate catalyst pool consists of candidate samples that have not yet obtained a true target catalytic performance label but have completed the calculation of catalyst descriptor vectors. The initial catalyst dataset contains catalyst descriptor vectors and corresponding target catalytic performance labels. The Gaussian process substitution model training module is used to train a Gaussian process substitution model based on the current dataset to obtain the predicted mean and predicted variance of catalytic performance. The causal graph learning and edge confidence estimation module is used to learn the causal structure of the current dataset using the PC algorithm to obtain the causal graph and the edge confidence of each edge in the causal graph; The causal intervention value calculation module is used to calculate the causal intervention value of each candidate sample based on the confidence of the updated front side, the confidence of the updated back side obtained by using the GP prediction mean as a pseudo-observation, and the uncertain edge weight. The expected improvement calculation module is used to calculate the expected improvement relative to the current best observation by minimizing the predicted mean and predicted standard deviation based on the Gaussian process substitution model and the current best observation. The adaptive weight calculation module is used to calculate adaptive weights based on the current iteration number and the maximum iteration number. The joint acquisition function construction module linearly combines the causal intervention value and the expected improvement into a joint acquisition function based on the adaptive weights. The candidate selection and label acquisition module is used to select the optimal candidate sample and obtain the true label based on the joint acquisition function. And an iteration control module, which controls each module to perform multiple iterations in sequence until the maximum number of iterations is reached.

[0017] This invention offers the following advantages: It improves the quality of causal graph structure recovery while actively selecting samples. It identifies descriptors with high explanatory value for target catalytic performance and reduces reliance on spurious descriptors that may be affected by confounding factors; for example, in heuristic analysis for oxygen evolution reaction (OER) catalyst discovery, this capability can be further used to assist in identifying candidate descriptors related to the catalytic mechanism. Therefore, this invention not only optimizes prediction performance during automated iterative experiments but also clarifies the iterative causal structure of catalyst discovery and improves the quality of causal graph structure recovery during the iterative process. Furthermore, the adaptive weighting strategy employed in this invention balances causal structure exploration and target performance optimization under different experimental budgets. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of a causal perception-based active learning method for catalyst discovery according to the present invention.

[0019] Figure 2 This is a system architecture block diagram of the present invention.

[0020] Figure 3 This is a causal structure recovery curve diagram of each method on the synthesis baseline in Embodiment 1 of the present invention.

[0021] Figure 4 This is a bar chart comparing the ablation experiments in Example 1 of the present invention.

[0022] Figure 5 This is a schematic diagram of the causal graph inferred by CIAL on the Materials Project OER dataset in Embodiment 2 of the present invention.

[0023] Figure 6 This is the adaptive weight evolution diagram in Embodiment 3 of the present invention. Detailed Implementation

[0024] The following detailed description of the embodiments, with reference to the accompanying drawings, will further illustrate the specific implementation of the present invention, in order to help those skilled in the art to have a more complete, accurate, and in-depth understanding of the inventive concept and technical solution of the present invention.

[0025] Example 1.

[0026] like Figures 1-6As shown, the present invention provides a causal perception active learning method for catalyst discovery, comprising the following steps.

[0027] S1. Data initialization and candidate pool construction: Obtain the candidate catalyst pool and the initial catalyst label dataset.

[0028] In this step, the candidate catalyst pool consists of candidate catalyst samples that have not yet obtained a true target catalytic performance label. The candidate samples are derived from historical material database screening, composition space enumeration, structure generation rules and / or high-throughput pre-screening results, and each candidate sample has completed catalyst descriptor vector calculation.

[0029] The initial catalyst label dataset is formed by density functional theory (DFT) calculation or experimental labeling of randomly sampled samples and / or preset seed samples from the candidate catalyst pool, and includes catalyst descriptor vectors and corresponding target catalytic performance labels.

[0030] The catalyst descriptor vector includes electronic structure features, morphological features and / or composition features; the target catalytic performance label is the catalytic performance index to be optimized, and the catalytic performance index can be formation energy, adsorption energy, overpotential or other target properties characterizing catalytic activity, stability or selectivity.

[0031] Specifically, the candidate catalyst pool is denoted as Its relationship with the current dataset The relationship is: The samples already have labels indicating the actual target catalytic performance. The candidate samples in the pool have not yet obtained true labels but have completed descriptor calculation; each round starts from... Select a candidate sample for DFT calculation or experimental labeling, and add the sample and its true label to the database. At the same time from Remove the sample from the list.

[0032] S2. Training of Gaussian Process Alternate Model: Based on the current dataset, train the Gaussian Process (GP) alternative model to obtain the predicted mean and predicted variance of catalytic performance.

[0033] In this step, the Gaussian process substitution model uses a radial basis function (RBF) kernel, and the hyperparameters are optimized by maximizing the marginal likelihood.

[0034] Specifically, the Gaussian process substitution model employs the RBF kernel function. When the dataset is small, a fixed default length scale is used to avoid instability in hyperparameter estimation; when the data volume is sufficient, the kernel hyperparameters are optimized by maximizing the marginal likelihood. Given a training input matrix X and a label vector y, the Gaussian process substitution model provides a predicted mean for candidate samples x. and prediction variance Where k is the kernel function and K is the kernel matrix. I represents the variance of the observed noise, and I is the identity matrix.

[0035] S3. Causal Graph Learning and Edge Confidence Estimation: The PC algorithm is used to learn the causal structure of the current dataset to obtain the causal graph and the edge confidence of each edge in the causal graph.

[0036] In this step, the PC algorithm is used to complete the causal graph in the form of a Completed Partially Directed Acyclic Graph (CPAG). In each iteration, the causal graph is learned based on the current dataset, and the CPDAG form is output.

[0037] In each iteration, the edge confidence of each edge in the causal graph is obtained through Bootstrap resampling (sampling with replacement). Specifically, the Bootstrap resampling is performed 100 times, and the edge confidence is obtained by statistically analyzing the frequency and directional consistency of each edge in the subsample causal graph; for each possible edge (i,j), the proportion of edges identified as existing and having the same direction in the subsample is calculated to obtain the edge confidence before the update. For edges in a CPDAG whose direction is undetermined but which have adjacency relationships, they are denoted as structurally uncertain edges and included in the calculation of causal intervention value.

[0038] S4. Calculation of Causal Intervention Value (IV): Based on the confidence of the updated front side, the confidence of the updated back side obtained by using the GP prediction mean as a pseudo-observation, and the weight of the uncertain side, the causal intervention value of each candidate sample is calculated.

[0039] In this step, the predicted mean obtained in step S2 is used as the confidence level of the updated side and the weight of the uncertain side obtained from the pseudo-observation. Combined with the confidence level of the updated side, the structural change proxy index (causal intervention value) inspired by the idea of ​​structural Hamming distance is calculated and used as the causal intervention value.

[0040] Specifically, the causal intervention value approximation adopts a single-step proxy update, where a Gaussian process is used to replace the model's predicted mean for a candidate sample x. As a pseudo-observation label, the simulation will Add to the current dataset to form a temporary dataset. The Bootstrap PC algorithm was then rerun on this temporary dataset to obtain the updated backside confidence. .

[0041] Calculate the change in edge confidence The formula is Calculate the uncertainty weights The formula is The causal intervention value is obtained by weighting and summing the uncertainty weights of each uncertain edge with the corresponding edge confidence changes. The formula is: The causal intervention value metric measures the potential clarifying effect of candidate samples on structurally uncertain edges. It is a structural change proxy metric based on changes in edge confidence, rather than a direct calculation of the standard SHD.

[0042] S5. Expected Improvement (EI) Calculation: For each candidate sample in the candidate pool, based on the predicted mean and predicted standard deviation of the Gaussian process substitution model, and the current best observation, calculate the expected improvement relative to the current best observation by minimizing the objective.

[0043] In this step, the target catalytic performance label is a minimum target value that is preferred when the value is as low as possible, based on the predicted mean. and the predicted standard deviation Define the current best observation value as Standardized variables are Expected to be improved to Where Phi is the standard normal distribution function and phi is the standard normal probability density function; when the prediction standard deviation When =0, the expected improvement is Take 0.

[0044] S6. Adaptive weight calculation: Calculate adaptive weights based on the current iteration number and the maximum iteration number.

[0045] In this step, the adaptive weight calculation formula is: Where N is the current iteration number, counted starting from 1. Maximum number of iterations; early stage Larger, prioritized causal exploration makes the joint acquisition function more biased towards IVs to clarify the causal structure; later Smaller values ​​are prioritized for prediction utilization, making the joint acquisition function more biased towards EI to select candidate samples that are more likely to improve target performance.

[0046] S7. Construction of the joint acquisition function: Based on the adaptive weights, the causal intervention value and the expected improvement are linearly combined into a joint acquisition function.

[0047] The joint acquisition function is ,in For adaptive weights, For the value of causal intervention, For improvement.

[0048] S8. Candidate Selection and Label Acquisition: Select the optimal candidate sample and obtain the true label based on the joint acquisition function.

[0049] This step selects the optimal candidate sample from the candidate catalyst pool based on the joint acquisition function, performs density functional theory calculations on the optimal candidate sample to obtain the real target catalytic performance label, and adds it to the dataset.

[0050] S9. Iterate until convergence: Repeat steps S2 to S8 until the maximum number of iterations is reached, and output the final dataset and the corresponding causal graph.

[0051] This invention also provides a causal perception active learning system for catalyst discovery, including a data initialization module, a Gaussian process substitution model training module, a causal graph learning and edge confidence estimation module, a causal intervention value calculation module, an expected improvement calculation module, an adaptive weight calculation module, a joint acquisition function construction module, a candidate selection and label acquisition module, and an iteration control module.

[0052] The specific technical descriptions of each of the above modules are as follows: The data initialization module is used to obtain a candidate catalyst pool and an initial catalyst dataset. The candidate catalyst pool consists of candidate samples that have not yet obtained a true target catalytic performance label but have completed the calculation of catalyst descriptor vectors. The initial catalyst dataset contains catalyst descriptor vectors and corresponding target catalytic performance labels.

[0053] The Gaussian process substitution model training module is used to train a Gaussian process substitution model based on the current dataset to obtain the predicted mean and predicted variance of catalytic performance.

[0054] The causal graph learning and edge confidence estimation module is used to learn the causal structure of the current dataset using the PC algorithm to obtain the causal graph and the edge confidence of each edge in the causal graph.

[0055] The causal intervention value calculation module is used to calculate the causal intervention value of each candidate sample based on the confidence of the updated edge, the confidence of the updated edge obtained by using the GP prediction mean as a pseudo-observation, and the weight of the uncertain edge.

[0056] The expected improvement calculation module is used to calculate the expected improvement relative to the current best observation, based on the predicted mean and predicted standard deviation of the Gaussian process substitution model and the current best observation, with the goal of minimizing the expected improvement.

[0057] The adaptive weight calculation module is used to calculate adaptive weights based on the current iteration number and the maximum iteration number.

[0058] The joint acquisition function construction module linearly combines the causal intervention value and the expected improvement into a joint acquisition function based on the adaptive weights.

[0059] The candidate selection and label acquisition module is used to select the optimal candidate sample and obtain the real label based on the joint acquisition function.

[0060] The iteration control module is used to control each module to perform multiple iterations in sequence until the maximum number of iterations is reached.

[0061] The effectiveness of the above method is verified through specific embodiments and experiments, as follows: Example 1: Synthetic benchmark verification.

[0062] Experimental setup: Generate a dataset of synthetic catalysts containing 10 variables, with a maximum number of iterations N. max =16, with five different randomized seeds used for repeated experiments. Comparison methods included random sampling, uncertainty sampling, expected improvement in baseline (EI), causal intervention value only (CIAL-IV-only), and expected improvement only (CIAL-EI-only).

[0063] like Figure 3 As shown, the evaluation metrics include SHD and Edge F1. SHD, or Structural Hamming Distance, measures the number of edges that need to be added, deleted, or reoriented between the learned causal graph and the real causal graph; a smaller value indicates a smaller structural difference. Edge F1 is the harmonic mean of edge recognition accuracy and recall; a larger value indicates better edge recognition performance. Experimental results show that CIAL-full achieves SHD 3.0±0.6 and Edge F1 0.87±0.05; EI baseline achieves SHD 12.6±3.3 and Edge F1 0.25±0.05; random sampling achieves SHD 15.4±3.3 and Edge F1 0.21±0.03; and uncertain sampling achieves SHD 12.8±2.8 and Edge F1 0.25±0.04. Therefore, under the conditions of the embodiments, this invention can more effectively recover the causal structure between catalyst descriptors compared to strategies that only consider performance improvement or random selection.

[0064] like Figure 4As shown, the ablation results indicate that CIAL-IV-only ( =1) achieved SHD 7.4±0.5, Edge F10.68±0.05; CIAL-EI-only ( =0) achieved SHD 9.4±0.9 and Edge F1 0.56±0.05. In Adaptive The strategy's SHD / F1 ratios were 8.2 / 0.72, 6.5 / 0.82, and 4.8 / 0.89, respectively, all outperforming the fixed-weight strategy (fixed weight). =0.5 corresponds to 9.5 / 0.65, 7.8 / 0.75, and 5.9 / 0.83; fixed =1.0 corresponds to 9.8 / 0.61, 7.5 / 0.70, and 6.2 / 0.80; fixed. =0.0 corresponds to 12.3 / 0.50, 10.0 / 0.55, and 8.1 / 0.64. Therefore, adaptively combining causal structure exploration with performance improvement can avoid the problems of overexploration by a single IV strategy or neglect of causal structure by a single EI strategy.

[0065] Computational overhead: In the unified experimental workflow of the synthetic benchmark, the runtime per round of random sampling, uncertainty sampling, and EI baseline is approximately 0.06–0.07 seconds; CIAL-IV-only is approximately 1.94 ± 0.20 seconds; CIAL-EI-only is approximately 1.60 ± 0.23 seconds; and CIAL-full is approximately 2.63 ± 0.50 seconds. Therefore, given the data scale of the examples, the computation time for each round of candidate evaluation in this invention is in the second range, which is acceptable in DFT calculation or experimental annotation workflows. Example 2: Validation of transition metal oxide benchmarks using the Materials Project database.

[0066] Dataset: The Materials Project transition metal oxide benchmark dataset was used, containing 41,758 filtered oxide entries. Filtering criteria included transition metal oxides, 2 to 5 elements, and energies higher than the convex hull and less than 0.2 eV. The target label was `formation_energy_per_atom`, with lower values ​​preferred. Core descriptors included `band_gap`, `efermi`, `density`, `volume`, `nsites`, `energy_above_hull`, and `mean_electronegativity`. Further compositional descriptors for Fe and Co were added in some follow-up experiments.

[0067] Performance of the Gaussian Process (GP) Alternative Model: On the Materials Project benchmark, the Gaussian process alternative model with the RBF kernel function fits the formation_energy_per_atom as follows: (Predicted mean) and The R² of the predicted standard deviation is 0.994, and the mean absolute error (MAE) is 23.1 meV / atom. Therefore, the GP model can provide reliable predicted mean and predicted standard deviation on the baseline descriptor, meeting the need for improved computational uncertainty quantification. Experimental setup: A cross-validation protocol based on the inference reference graph was adopted. In a real-world data scenario without a true causal graph, graph recovery F1 was used as the comparison metric. The main baseline used 7 core descriptors with 5 random seeds for repetition. Furthermore, NOTEARS (Non-combinatorial Optimization via Trace Exponential and Augmented lagRangian for Structure learning) was used as the causal structure learning backend, and cross-validation was performed under 10-seed and 20-seed conditions.

[0068] Experimental results: such as Figure 5 As shown, on the main Materials Project benchmark (7 core descriptors, 5 seeds), CIAL's F1 score is 0.422±0.078, EI's is 0.445±0.080, and random sampling's is 0.361±0.116, indicating that the present invention is competitive with EI under this general descriptor setting. Under the NOTEARS cross-partitioning protocol, with 10 seeds, CIAL's F1 score is 0.433±0.088, higher than EI's 0.327±0.091 (p=0.0216); with 20 seeds, CIAL's F1 score is 0.408±0.103, higher than EI's 0.340±0.107 (p=0.0520, significant boundary).

[0069] Therefore, under the conditions of real material screening data, the present invention can maintain graph recovery performance similar to standard EI on the primary reference, and exhibits higher relative inferred graph recovery capability under the specific NOTEARS cross-partitioning protocol. Since there is no real causal graph in this real data scenario, the above F1 is a comparison index relative to the inferred reference graph, used to illustrate the applicability of the present invention in real material screening protocols, rather than a direct verification of absolute causal correctness.

[0070] Cross-dataset supplementary validation: In exploratory validation on the OC20 Catalyst dataset (5 seeds, NOTEARS cross-split), the F1 score for CIAL was 0.502±0.100, and the EI was 0.438±0.036. The differences were in the same direction but not statistically significant (p=0.2826). This result is only used as supplementary context across datasets and does not constitute the main validation basis.

[0071] Example 3: Adaptive weight evolution analysis.

[0072] like Figure 6 As shown, during the 16 rounds of active learning on synthetic data (random data with known causal relationships inspired by the Materials Project database), According to the adaptive weight calculation formula The preset linear decay law gradually decreases. In the early iterations, Larger (e.g., the first 5 rounds) >0.6), making the joint acquisition function more biased towards causal intervention value (IV) to prioritize clarifying uncertain edges in the causal structure; as the iteration progresses, Gradually reduce (last 5 rounds) <0.3), the focus of the joint acquisition function shifts to the expectation of improving EI, so as to prioritize candidate samples that are more likely to improve the target performance; around the 9th round, the utility of IV and EI tends to be balanced.

[0073] Therefore, this invention can achieve a smooth transition from causal exploration to predictive utilization within the same active learning process, thereby balancing structure recognition and performance optimization within a limited experimental budget.

[0074] Table 1 shows the abbreviations of the technical terms used in the above technical solutions.

[0075] Table 1. Abbreviations for Professional and Technical Terms

[0076] The present invention improves the accuracy of causal structure recovery under the conditions of the embodiments: on the synthetic benchmark, CIAL exhibits a lower Structural Hamming Distance (SHD) and a higher Edge F1 compared to the EI baseline, thus demonstrating that the present invention can improve the quality of causal graph structure recovery while actively selecting samples.

[0077] Ablation experiments demonstrate that the adaptive weighting strategy can balance causal structure exploration and target performance optimization under different experimental budgets. Furthermore, this invention offers the advantages of controllable computational overhead and no need for manual intervention. Controllable computational overhead: Under the data scale of the examples, the computation time for each round of candidate evaluation is in the second range, which can be used in conjunction with DFT calculations or experimental annotation processes; fully automated and requiring no manual intervention, it is suitable for large-scale high-throughput screening.

[0078] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A causal perception-based active learning method for catalyst discovery, characterized in that, Includes the following steps: S1. Data initialization and candidate pool construction: Obtain the candidate catalyst pool and the initial catalyst label dataset; S2. Training of Gaussian process substitution model: Train the Gaussian process substitution model based on the current dataset to obtain the predicted mean and predicted variance of catalytic performance; S3. Causal Graph Learning and Edge Confidence Estimation: The PC algorithm is used to learn the causal structure of the current dataset to obtain the causal graph and the edge confidence of each edge in the causal graph. S4. Causal intervention value calculation: Based on the confidence of the updated front side, the confidence of the updated back side obtained by using the prediction mean of the Gaussian process replacement model as a pseudo-observation, and the weight of the uncertain side, the causal intervention value of each candidate sample is calculated. S5. Expected Improvement Calculation: Based on the predicted mean and predicted standard deviation of the Gaussian process substitution model, and the current best observation, calculate the expected improvement relative to the current best observation by minimizing the objective. S6. Adaptive weight calculation: Calculate adaptive weights based on the current iteration number and the maximum iteration number; S7. Construction of Joint Acquisition Function: Based on the adaptive weights, the causal intervention value and the expected improvement are linearly combined into a joint acquisition function; S8. Candidate Selection and Label Acquisition: Select the optimal candidate sample and obtain the true label based on the joint acquisition function; S9. Iterate until convergence: Repeat steps S2 to S8 until the maximum number of iterations is reached, and output the final dataset and the corresponding causal graph.

2. The causal perception-based active learning method for catalyst discovery according to claim 1, characterized in that, In step S2, the Gaussian process substitution model uses a radial basis function kernel, and the hyperparameters are optimized by maximizing the marginal likelihood. When the dataset size is small, a fixed default length scale is used to avoid unstable hyperparameter estimation. When the amount of data is sufficient, the kernel hyperparameters are optimized by maximizing the marginal likelihood.

3. The causal perception-based active learning method for catalyst discovery according to claim 2, characterized in that, In step S3, the causal graph is represented as a partially directed acyclic graph. In each iteration, the edge confidence of each edge in the causal graph is obtained through Bootstrap resampling. The edge confidence is obtained by statistically analyzing the frequency and directional consistency of each edge in the subsample causal graph. For each possible edge (i,j), the proportion of edges identified as existing and having the same direction in the subsample is calculated to obtain the edge confidence before the update. In a CPDAG, edges with undetermined directions but adjacency relationships are denoted as structurally uncertain edges.

4. The causal perception-based active learning method for catalyst discovery according to claim 3, characterized in that, In step S4, the predicted mean of the replacement model is used with a Gaussian process. As a pseudo-observation label, the simulation will Add the current dataset to form a temporary dataset, and rerun the Bootstrap PC algorithm on this temporary dataset to obtain the updated confidence scores. ; Calculate the change in edge confidence The formula is Calculate the uncertainty weights The formula is The causal intervention value is obtained by weighting and summing the uncertainty weights of each uncertain edge with the corresponding edge confidence changes. It is used to measure the potential clarifying effect of candidate samples on structurally uncertain edges.

5. The causal perception-based active learning method for catalyst discovery according to claim 4, characterized in that, In step S5, based on the predicted mean and the predicted standard deviation Define the current best observation value as Standardized variables are Expected to be improved to Where Phi is the standard normal distribution function and phi is the standard normal probability density function; when the prediction standard deviation When =0, the expected improvement is Take 0.

6. A causal sensing active learning method for catalyst discovery according to claim 5, characterized in that, In step S6, the adaptive weight calculation formula is: Where N is the current iteration number, counted starting from 1. Maximum number of iterations; early stage Larger, prioritized causal exploration makes the joint acquisition function more biased towards IVs to clarify the causal structure; later Smaller values ​​are prioritized for prediction utilization, making the joint acquisition function more biased towards EI to select candidate samples that are more likely to improve target performance.

7. The causal perception-based active learning method for catalyst discovery according to claim 6, characterized in that, In step S7, the joint acquisition function is: ,in For adaptive weights, For the value of causal intervention, For improvement.

8. The causal sensing active learning method for catalyst discovery according to claim 1, characterized in that, In step S1, the candidate catalyst pool consists of candidate catalyst samples that have not yet obtained a true target catalytic performance label. The candidate samples are obtained from historical material database screening, composition space enumeration, structure generation rules and / or high-throughput pre-screening results, and each candidate sample has completed the catalyst descriptor vector calculation. The initial catalyst label dataset is formed by random sampling samples and / or preset seed samples in the candidate catalyst pool after density functional theory calculation or experimental labeling, and includes catalyst descriptor vectors and corresponding target catalytic performance labels.

9. The causal perception-based active learning method for catalyst discovery according to claim 2, characterized in that, In step S2, the Gaussian process substitution model provides a predicted mean for the candidate sample x. and prediction variance Where k is the kernel function and K is the kernel matrix. I represents the variance of the observed noise, and is the identity matrix.

10. A causal sensing active learning system for catalyst discovery, characterized in that, include: The data initialization module is used to obtain the candidate catalyst pool and the initial catalyst dataset. The candidate catalyst pool consists of candidate samples that have not yet obtained true target catalytic performance labels but have completed catalyst descriptor vector calculations. The initial catalyst dataset contains catalyst descriptor vectors and corresponding target catalytic performance labels. The Gaussian process substitution model training module is used to train a Gaussian process substitution model based on the current dataset to obtain the predicted mean and predicted variance of catalytic performance. The causal graph learning and edge confidence estimation module is used to learn the causal structure of the current dataset using the PC algorithm to obtain the causal graph and the edge confidence of each edge in the causal graph; The causal intervention value calculation module is used to calculate the causal intervention value of each candidate sample based on the confidence of the updated front side, the confidence of the updated back side obtained by using the GP prediction mean as a pseudo-observation, and the uncertain edge weight. The expected improvement calculation module is used to calculate the expected improvement relative to the current best observation by minimizing the predicted mean and predicted standard deviation based on the Gaussian process substitution model and the current best observation. The adaptive weight calculation module is used to calculate adaptive weights based on the current iteration number and the maximum iteration number. The joint acquisition function construction module linearly combines the causal intervention value and the expected improvement into a joint acquisition function based on the adaptive weights. The candidate selection and label acquisition module is used to select the optimal candidate sample and obtain the true label based on the joint acquisition function. And an iteration control module, which controls each module to perform multiple iterations in sequence until the maximum number of iterations is reached.

Images